ORION

Pre-requisites

Matlab

Download Tensor and Poblano Toolbox and add it to Matlab Path

Python

Scipy
Sklearn
Numpy
Matplotlib
Tensorly

If you are using Anaconda for Python Packages, you can use the following commands:

# Create a new environment, so that you don't mess up your exisiting environment
conda create --name <your new env name>
conda activate <your new env name>
conda install scikit-learn
conda install -c tensorly tensorly
conda install -c conda-forge matplotlib

Datasets

How to Run

Git clone or download this code
Run init.m in matlab to create required folders like dataset, results, tensorDataset.
Download dataset(s) from the above provided links into the dataset folder.

In runDataset.m set below variables according to the dataset you are using. For example if you are running it for IndianPines dataset:

   datasetFname = 'dataset/Indian_pines_corrected.mat'
   datasetGt = 'dataset/Indian_pines_gt.mat'
   outFile= 'IndianPines'
   % IMPORTANT: Change X and Y according to variable stored in the .mat(dataset) file
   X = data.indian_pines_corrected;
   Y = gt. indian_pines_gt;
   
   testSize = 0.2
   % Number of datasets to be created
   numData = 10
   % Tensor decompostion rank
   ranks = [1000, 2000]

After running runDataset.m using Matlab, it will create .mat files in tensorDataset/8020/IndianPines based on outFile and testSize variable (in above example testSize was 0.2, so 80-20 split).

Now to run ORION method, navigate to python folder and set following variables in orion.py file.

    dataPath = '../tensorDataset/8020/IndianPines/'

After running the orion.py file, it will generate results(figures and .mat files) in results/orion/8020/IndianPines/. Path of the result depends on the dataset being used and train-test split(in above example testSize was 0.2, so 80-20 split).

We have also provided the code for baselines used in our paper, to run Linear, Polynomial and RBF SVM set follwing variables in baselines.py and to run Multi Layer perceptron set the same following variables in mlpBaseline.py file:

   dataX = loadmat('../dataset/Indian_pines_corrected.mat')
   dataY = loadmat('../dataset/Indian_pines_gt.mat')
   Xog = dataX['indian_pines_corrected'] # 3D object
   Y2d = dataY['indian_pines_gt']        # 2D object
   folderName = 'IndianPines' # Make sure this is correct. Result folders will be created based on this.
   
   # number of runs
   runs = 10
   
   testSize = 0.2

7.(Optional) Our code uses GridSearchCV from sklearn for hyperparameter tuning, to make it run faster(in parallel) you can set njobs variable in trainModelSVM and trainNN in models.py according to your system configuration. For details refer to this link